Contrastive representation learning for self-supervised deception detection in edge LLMs
Feng An, Wenyin Tao · 2026
Read Contrastive representation learning for self-supervised deception detection in edge LLMs by Feng An, Wenyin Tao (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/plos-pone-0354894.txt.
Description
Feng An and Wenyin Tao tackle a hard safety problem: spotting when a large language model reasons in deceptive ways. Older monitors label whole answers as safe or deceptive and lean on heavyweight teacher models. Their 2026 paper instead trains a tiny monitor that reads the model's hidden chain-of-thought states.
The monitor maps those states into a space where honest and deceptive reasoning sit apart. Training pulls similar cases together and pushes unlike ones away, so gradual shifts toward deception become visible. On a sycophancy test set, the contrastive monitor beat a standard classifier at flagging deceptive tendency. An and Tao argue that watching reasoning geometry offers a lighter path to self-supervised safety checks.
Catalog record
| CREATOR | Feng An, Wenyin Tao |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 43545 |
| LICENSE | CC_BY |
| RIGHTS | CC BY |
| SOURCE | PLOS |